Our goal is to build a super-large-scale deep neural network capable of matching the complexity of the human brain, thereby achieving a quantum leap in artificial intelligence. To achieve this, we have set three research objectives.
First, parallel processing by multiple general-purpose computers to train large-scale models efficiently in terms of cost and time.
Second, self-learning leveraging large amounts of unlabeled raw data for efficient and scalable training.
Third, speaker anonymization that removes sensitive personal information, such as speaker identity, while preserving linguistic information as much as possible, thus preventing privacy invasion during large-scale data collection.
By achieving these three objectives, we can build large-scale deep neural networks cost-effectively while ensuring privacy protection.
Our decades of experience in speech processing technologies, such as speech recognition, speaker identification, and sound source localization, can play a key role in validating the effectiveness of building large-scale deep neural networks.